Surprising news appeared - an artist spent a 100
hours creating something truly incredible for a book cover image…unfortunately, only to be accused
having used an AI. However, when I looked at this, my first thought was that this looks way too
good for an AI image in 2023. Why is that? Well, look. There is so much detail in this image,
no state of the art text to image AI system can match it. The resolution of this image, the
number of pixels are just too high. Until now. Dear Fellow Scholars, this is Two Minute
Papers with Dr. Károly Zsolnai-Fehér. Now, previous techniques already exist to
increase the resolution of input images, or even videos. This is what we
call upscaling or super resolution, and it is starting to appear in video games, and
in some cases, it already runs in real time. So, why not apply it to these AI-made photos,
that should be easier than video, right? Well, not quite. You see, super resolution means
that we introduce new information into an image. And this new information is unfortunately, not
always what should be there. For instance, text can easily get mangled in video games. If we have
a stationary photo that we would like to frame, it better not have artifacts like this, right?
So, is all hope lost for high-resolution images? That was what I thought, until
I saw the news from Midjourney, a popular text to image AI system, which can
now create incredibly detailed, high-resolution images with a recent update. And the results
are truly insane. How does it work? Well, you first generate a lower resolution image,
these are pretty good, however, when we start zooming in, we start to see the limits of this
resolution. But then, hold on to your papers, because we can ask it to perform 4x super
resolution, and then, this happens. There is just so much more detail in these images, I am
completely stunned. Wow. This seems like years and years of progress, and it just dropped a few
days ago in an update like nothing had happened. Now, other super resolution or image
upscaling systems also exist. In fact, they are in abundance today. We have
so many options. However, this kind of quality is beyond what I have seen from other
solutions out there. According to other users, it is perhaps matched only by the Gigapixel
AI system, I gave it a brief try and did not have too much luck with it so far. It is also a
standalone application, which is not necessary when it is built right into the text to image AI.
For me, so far, this workflow has been unmatched. And it just delivers over and over again. My opinion is that these techniques
are deceivingly useful. You see, you may often have an image that is a
regular photograph or an AI-generated work, where you love the image, but the framing is not
quite perfect, and you would only like to use one small part of it. However, when you do that,
you take the parts that you like, and suddenly, there are just not enough details for this
image to stand on its own. And in these cases, these super resolution techniques can
make or break that image. Loving it. Now, not even this new technique is perfect, not
even close. In some cases, excessive smoothing happens, which is quite curious. Look. You would
think that this is the low resolution image, and this is the upscaled, but that is
not the case. Due to this smoothing, details are not only not added to the image,
but some can even disappear in the process! But, in return, the overall image and the
edges of the objects are way sharper, so there are cases where this is a
trade off that we have to consider. And, my goodness, through the
power of AI research, today, everyone can experience the amazing feeling of
their artistic ideas coming to life. There are already systems that even work on the phone
in your pocket. Wow. And to think that text to video also exists, just imagine that soon,
this is what will happen to videos too. Soon, everyone will be able to become a movie
director, first, for at a small cost, and then, over time, all this will be free
for all of us. What a time to be alive!
